Meta
AI feed and video recommendation (large-scale models / LLMs)
Meta's AI recommendation models added +5% time spent on Facebook and +6% on Instagram in Q2 2025, then +7% views of organic feed and video posts in Q4 2025.
Key points
- Rearchitecture of feed and video ranking and recommendation systems with AI.
- Proprietary large-scale ranking models, with LLM-style approaches.
- +5% time spent on Facebook and +6% on Instagram in Q2 2025.
- Evidence A, confirmed status, gains reconfirmed quarter after quarter.
Objective
Increase time spent and engagement on Facebook, Instagram, and Threads by showing each user more relevant content, which directly feeds ad inventory and revenue.
The deployment
Meta rearchitected its ranking and recommendation systems with larger-scale models, including LLM-style approaches applied to feed and video ranking. The goal is to better predict what each user will find interesting. Gains are measured in points of time spent and views, and reported quarter by quarter in earnings.
Results Proof A
Figures stated by Mark Zuckerberg on the Q2 2025 earnings call (reported by the press) and repeated and expanded in an official Meta communication for Q4 2025. Results from a public company, explicitly attributed to improvements in the recommendation systems.
How it works
Documented architectureThe stack in detail
- outil Modeles de ranking proprietaires Meta Large-scale deep learning recommendation that ranks candidate feed and Reels content; Meta's proprietary ML, not sold commercially.
- llm Architectures type LLM appliquees au ranking LLM-style approaches used for feed and video ranking according to Meta; the exact model is not publicly named.
- infra Infrastructure ML interne Meta Real-time serving of scoring at each feed load and training on interaction logs.
How it runs, concretely
For ops teams-
1Interaction collection site_app / data team
Views, time spent, likes, shares, and comments are logged per user and per content item.
-
2Feed ranking AI
The large-scale models rank candidate content to maximize predicted relevance.
-
3Personalized rendering site_app
Each user sees a feed and Reels ordered for them on Facebook, Instagram, or Threads.
-
4Measurement and iteration data team
Gains in time spent and views are measured and reported; models are iterated quarterly.
Time spent and interactions per user, which serve as the optimization objective and the measure. Without this continuous signal, the ranking models can neither train nor prove their gain.
How your customers perceive this type of use
Sourced studiesLe paradoxe est documente des deux cotes : 71% des consommateurs attendent des interactions personnalisees et 76% sont frustres quand elles manquent (McKinsey, 2021), mais 75% declarent ne pas acheter aupres d'organisations auxquelles ils ne confient pas leurs donnees (Cisco, 2024). La « creepy line » est localisee : messages recus quelques secondes apres une recherche et suivi de localisation sont les pratiques qui mettent le plus mal a l'aise (Periscope by McKinsey, 2019).
Acceptance conditions
- La confiance dans le traitement des donnees precede l'achat : 75% ne achetent pas sans elle (Cisco 2024)
- Un cadre legal protecteur rassure : 59% des consommateurs disent que des lois fortes sur la vie privee les rendent plus a l'aise pour partager des informations dans des applications IA (Cisco 2024)
- La personnalisation elle-meme est attendue quand elle est consentie : environ la moitie des consommateurs (US 55%, UK 52%) disent s'inscrire souvent ou parfois a des services personnalises (Periscope by McKinsey 2019)
Red lines
- Le message declenche quelques secondes apres une recherche ou un achat : deuxieme ou troisieme cause de malaise selon les pays (Periscope by McKinsey 2019)
- Le suivi de localisation percu comme de la surveillance : 40% de malaise en Allemagne et au Royaume-Uni (Periscope by McKinsey 2019)
- Le mesusage des donnees personnelles par l'IA, devenu la premiere inquietude des consommateurs, a 53% et en hausse (Qualtrics 2025)
Sources: McKinsey & Company 2021 · Periscope by McKinsey 2019 · Cisco 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- large-scale per-user interaction logs
- indexed content catalog
- explicit and implicit engagement signals
Org prerequisites
- a ranking ML team
- massive real-time serving infrastructure
- quarterly experimentation capacity
Possible stack
- custom/in-house
- large-scale recommendation models
- LLM-style architectures for ranking
The plan, step by step
- Step 1Establish a reliable measure of time spent and views per user, and set up an A/B testing framework on the feed.Deliverable: Validated engagement dashboard and experimentation protocol.
- Step 2Build the interaction logging pipeline and a ranking baseline measured in production.Deliverable: Documented baseline with its reference metrics.
- Step 3Train heavier recommendation models on the interactions and evaluate them offline against the baseline.Deliverable: Candidate model beating the baseline in offline evaluation.
- Step 4Run the A/B test in production on a fraction of traffic and measure time spent and views.Deliverable: Experiment readout with a significant gain or a stop decision.
- Step 5Generalize the winning model and set up quarterly iterations.Deliverable: Time-spent and view gains tracked and reported quarter by quarter.
First step: Establish a reliable measure of time spent and an A/B testing framework before investing in heavier ranking models.
The same pattern, proven in other industries
Boots
machine learning scoring of offer-usage propensity, applied to mail and the app
Michaels
genAI personalization of marketing messaging (copy) at scale, tuned to the brand voice
Nike
personalized exclusive offers and access based on member engagement and history
Sources
- S1 Zuckerberg: AI increased the time spent on Facebook and Instagram in Q2 (TechCrunch, earnings Q2 2025) Established press archive pending
- S2 2026: AI Drives Performance (About Meta) Interested party archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.